AEROS: A Physics-Informed Neural Network Pilot for Real-Time Autonomous Mars Entry and Descent Control

This paper presents AEROS (Aerocapture and Entry with Reinforced Learning for Opti mization and Survivability), a hybrid neural guidance system for autonomous Mars atmospheric entry. AEROS combines a Deep Neural Network (DNN), a Physics-Informed Neural Network (PINN), and a Proximal Policy Optimization (PPO) reinforcement learning layer to generate real-time control actions during skip-entry and descent. The system interprets a 23-feature observation vector including altitude, velocity, angle of attack, bank angle, heat flux, dynamic pressure, and normal load, and produces Δ pitch, Δ yaw, a safe-to-descend flag, and a landing zone score. Aerothermodynamic loads are provided by a high-fidelity Direct Simulation Monte Carlo (DSMC) solver, SPARTA, and post-processed into a structured dataset for surrogate modeling. Simulation results in Mars-relevant entry conditions demonstrate that AEROS can reduce peak heat flux and dynamic pressure excursions, while maintaining corridor tracking and limiting reaction control system propellant usage, supporting the case for autonomous entry guidance in future Mars missions